My purpose is very different. I wanted to communicate with any machine I owned. Every machine has an Onion hidden service, and every machine can talk to Onions seamlessly.
What this allows is I can code against a [hash].onion, and know that the data goes where I want it. I can run Mosquitto (MQTT database) on one node, and other nodes can publish data to it. It matters not where they are, what networks they reside on, or if I get the "DynDNS, firewall holes, dynamic-static internal IP", and the rest of that junk set up right.
I also use Node-Red, and can use .onion addresses as valid services elsewhere. It allows me the ultimate network flexibility. I think of .onion addresses as being on a "Really Long Ethernet Hub" that only listens to the machine talked to.
EDIT:
> If you run an arbitrary protocol over Tor without any of that, it's much easier for the server to fingerprint the client.
Agreed. I control both endpoints.
I wrote this: https://trac.torproject.org/projects/tor/wiki/doc/LinuxDNSre...
I wanted to be able to send data to Tor onion endpoints within my code, mixed with regular IP or DNS addresses. This modification allows the Linux system resolver handle Onions as it would any other address.